ggsDing/ASLNet

代码与文中的描述有冲突

epoch8260 opened this issue · 15 comments

epoch8260 commented

且复现后达不到文中的效果，请提供训练的checkpoint

ggsDing commented

Hi,
There are two versions of the segmentation network. ED_FCN is a basic encoder-decoder FCN, while FCN_SR is the described FCN with shape refinement modules. I do not think there are conflicts with the descriptions in the paper.

epoch8260 commented
ggsDing commented

let me check whether I still have them on the servers.

epoch8260 commented
ggsDing commented

I just uploaded a checkpoint of the FCN+SR model (see link below). This is not the specific saved file associated with the results in the paper, but the accuracy should be close.
I will also check if the results can be reproduced with the released codes. Unfortunately, I am already graduated and do not have enough devices, so this could take time.
checkpoint link

ggsDing commented

Note that the evaluation codes in this repository are not very rigorous, as the prediction is performed on cropped images. It is suggested to first stitch the cropped masks, and then calculate the final accuracy.

epoch8260 commented
  • 👍1
jerrywyn commented

I just uploaded a checkpoint of the FCN+SR model (see link below). This is not the specific saved file associated with the results in the paper, but the accuracy should be close. I will also check if the results can be reproduced with the released codes. Unfortunately, I am already graduated and do not have enough devices, so this could take time. checkpoint link

hi, are you sure you have uploaded the right checkpoint file? It's not correspondent to the code you provided.
The error message is:
Missing key(s) in state_dict: "SR.0.weight".............
Unexpected key(s) in state_dict: "DC.0.weight"................

And, which dataset is used to train the checkpoint file?

Royal-kev commented

谢谢大佬，祝大佬毕业顺利工作顺利
…
---- 回复的原邮件 ---- | 发件人 | @.> | | 日期 | 2022年05月14日 00:49 | | 收件人 | @.> | | 抄送至 | @.@.> | | 主题 | Re: [ggsDing/ASLNet] 代码与文中的描述有冲突 (Issue #3) | I just uploaded a checkpoint of the FCN+SR model (see link below). This is not the specific saved file associated with the results in the paper, but the accuracy should be close. I will also check if the results can be reproduced with the released codes. Unfortunately I am currently graduated and do not have enough device, so this could take time. checkpoint — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

大佬，在复现过程中，遇到了一些问题，方便的话，可以交流一下嘛

ggsDing commented

I just uploaded a checkpoint of the FCN+SR model (see link below). This is not the specific saved file associated with the results in the paper, but the accuracy should be close. I will also check if the results can be reproduced with the released codes. Unfortunately, I am already graduated and do not have enough devices, so this could take time. checkpoint link

hi, are you sure you have uploaded the right checkpoint file? It's not correspondent to the code you provided. The error message is: Missing key(s) in state_dict: "SR.0.weight"............. Unexpected key(s) in state_dict: "DC.0.weight"................

And, which dataset is used to train the checkpoint file?

Please change the name of modules 'SR->DC' in the model file and the checkpoint will work. This was trained on the Inria Dataset.

jerrywyn commented

I just uploaded a checkpoint of the FCN+SR model (see link below). This is not the specific saved file associated with the results in the paper, but the accuracy should be close. I will also check if the results can be reproduced with the released codes. Unfortunately, I am already graduated and do not have enough devices, so this could take time. checkpoint link

hi, are you sure you have uploaded the right checkpoint file? It's not correspondent to the code you provided. The error message is: Missing key(s) in state_dict: "SR.0.weight"............. Unexpected key(s) in state_dict: "DC.0.weight"................
And, which dataset is used to train the checkpoint file?

Please change the name of modules 'SR->DC' in the model file and the checkpoint will work. This was trained on the DeepGlobe Dataset.

I tried to use your code to train on the Mas dataset and AIRS (WHU Building) dataset and found that discriminatior only degrades the performance, why is that? My configuration is crop_size=512, batch_size=4.

ggsDing commented

I just uploaded a checkpoint of the FCN+SR model (see link below). This is not the specific saved file associated with the results in the paper, but the accuracy should be close. I will also check if the results can be reproduced with the released codes. Unfortunately, I am already graduated and do not have enough devices, so this could take time. checkpoint link

hi, are you sure you have uploaded the right checkpoint file? It's not correspondent to the code you provided. The error message is: Missing key(s) in state_dict: "SR.0.weight"............. Unexpected key(s) in state_dict: "DC.0.weight"................
And, which dataset is used to train the checkpoint file?

Please change the name of modules 'SR->DC' in the model file and the checkpoint will work. This was trained on the DeepGlobe Dataset.

I tried to use your code to train on the Mas dataset and AIRS (WHU Building) dataset and found that discriminatior only degrades the performance, why is that? My configuration is crop_size=512, batch_size=4.

Hi, I ran the codes again and confirmed that steady improvements are made on the MAS dataset. Please check the uploaded training script for the MAS dataset in the main directory. It is recommended to remove 'conv3' in the discriminator before training, since the resolution of MAS is lower.